Descriptions:
Francesco Bonacci, CEO of Cua, presents at AI Engineer on what he calls Computer Use 2.0 — a shift from single-agent screen-takeover loops to parallelized, background-executing agents that never interrupt the user’s desktop. The centerpiece is Quad Driver, an open-source universal computer-use driver that works across macOS, Windows, and Linux by combining accessibility tree inspection with pixel-level fallback clicks, using undocumented Apple framework APIs to execute actions in the background without seizing screen control.
The team introduces KUBench, a benchmark suite now containing over 130 verifiable tasks across 42 environments and five platforms, co-developed with Snorkel AI to include electrical engineering tasks that simulate real circuits. Results are sobering: the top agent tested completed only 6 of 25 circuit tasks, all involving editing existing schematics — success rate on tasks starting from a blank schematic dropped to 0%. Across all tested models, no agent exceeded 30% reward on the full leaderboard.
On the positive side, swapping the computer-use tool in standard agents from the built-in option to Quad Driver improved pass rates from 62% to 80% on KUBench Basic tasks at 4K resolution, using 34% fewer tokens — primarily because Quad Driver focuses on the active window rather than the full desktop. Bonacci also introduces a multi-cursor paradigm: multiple specialized agents operating concurrently in parallel background sessions, each handling a different task without user interruption, which he frames as the natural next step beyond single-agent computer use.
📺 Source: AI Engineer · Published July 15, 2026
🏷️ Format: Keynote Launch
